Trained on HeavenOrangeMix but other anime models should also work
The model tends to produce close-ups, so I recommend adding close-up to negatives. Additionally, the default face expression isn't neutral and tends towards tokens like :3 or wavy mouth. Adding face expressions to your prompts or adding the above to your negative prompt should improve your generations. After I get some time to clean up the dataset and retrain the model, these problems should go away.
